Every statistical test – in fact, every comparison of every kind – that we do is based on some assumptions. At base, we assume that the data we are analyzing is not fabricated and that there were not egregious errors in how the data was collected.
Virtually all inferential statistics have some assumptions that have to be met. Think of them as the conditions we need to meet for our tests to work properly. Researchers check on these assumptions as they are preparing to conduct their analyses.
At the simplest level, many tests are applicable to only quantitative or qualitative data, but not both. However, there are other, more technical assumptions. Below we discuss some of the most common.
Independence of Observations
One assumption is about the independence of our observations. This means that each measure we collect is not influenced by the other measures we take. For example, if measuring how many buyers click on an ad, the assumption is that each buyer’s decision to click is made on their own. If one individual on a buying team were to click on the ad, go to a landing page, download an asset and send it to another, that would alter the behavior other the second individual — likely influencing them to not bother clicking on the ad. In this example, we may never know if the measures we have taken are completely independent, but we can assume that most browsers are not emailing each other the content.
If an industry influencer who frequently writes and speaks on the topic of measurement surveys her LinkedIn followers about their measurement practices, their responses may be biased by what they have read and heard from the influencer, and their responses might not represent the larger population of marketers. Researchers must always be on guard against such influences in research data.
In our research, while we do survey some of our customers, we also gather responses from a panel of B2B practitioners who are not affiliated with us. And, we compare how our customers respond compared to others. This approach helps ensure that our data reflects a broader perspective.
Normality
Many hypothesis tests assume that data are distributed normally. Many statistical tests simply cannot be applied to data that is not normally distributed. In some cases, there are accepted alternative tests for non-normally distributed data, as well as methods for transforming data so that statistical tests requiring normality can be applied.
Equality of Variance
In statistics, equality of variance means that two (or more) groups you want to compare have a similar range of values for the thing you want to measure. For example, if you want to compare the productivity of two teams of BDRs, it is important to understand if they vary across similar ranges of performance.
There are standard statistical tests to measure equality of variance. And there are statistical workarounds where the equality of variance assumption is not met.
If the BDRs on one team produce between 1 and 20 qualified opportunities per month, while the other team has between 50 and 200, standard statistical tests may need to be adjusted to account for the disparity in variability.
